RL makes AI math proofs insanely hard to read as correctness rewards trump human style
burny_tech · x · 2026-11-21
A notable model-behavior observation: reinforcement learning makes AI-generated math proofs "insanely hard to read" — RL is not constrained to the human distribution as much, and is driven primarily by correctness rewards. Proofs only need to be right, not human-readable, which is reshaping the legibility of AI math output.
More from Models
- MaxxBench cost-quality chart: Muse Ember 1.0 hits 63.5% at $0.18, crushing rivals on value — AIatMeta · 2026-11-21
- Thinking Machines releases Inkling 2 Mini: looped-transformer open model that runs on 20GB locally — mark_k · 2026-11-21
- Thinking Machines' Inkling-2-Mini trends on Hugging Face — thinkingmachines · 2026-11-21
- Inkling 2 Mini runs a local Mac assistant: 26.4B params, only 3.4B active — simonguozirui · 2026-11-21
- dottxt tool-call index: DeepSeek tops 100M daily calls, failure rates vary 0.04%–17% by provider — remilouf · 2026-11-21
- Deep dive: the claimed Claude-assisted Riemann proof uses theta lifts instead of derivatives, formalized in Lean — Frequent-Pen-9898 · 2026-11-21